The headline finding: the integrity case count on your dashboard is a measure of your own detection effort, your policy definitions and your reporting culture — in that order — before it is a measure of student behaviour. Change any of the three and the series moves without a single student changing what they do.
This matters now because integrity caseloads are being put into governance papers, business cases and press responses as though they were an incidence measure. They are not, and the difference is straightforward to state once, in writing, in a way that protects every later use of the number.
What is actually published, and what is not
Start with the negative result, because it saves a fortnight of searching. There is no recurring, cross-institutional series of academic integrity cases published as official statistics in the way enrolment, qualification and finance data are. National statistical agencies collect what institutions are required to return, and misconduct case counts are not among those returns in the markets this site covers.
What circulates instead falls into four categories, and they are not interchangeable:
| Source type | What it gives you | What it cannot support |
|---|---|---|
| Individual institutions’ annual integrity reports | A real internal series, with local definitions | Comparison with any other institution |
| Figures released under access-to-information requests | Snapshots across several institutions at once | A trend — the sample changes with each request |
| Vendor-published material | Product-relevant framing | Independent evidence for a business case |
| Self-report behaviour surveys | Prevalence estimates with a stated sample | Anything about cases, which are institutional acts |
The fourth row is the confusion that does the most damage. A survey saying some percentage of students did something is a statement about behaviour. A case count is a statement about what an institution detected, defined as reportable, and processed. They can move in opposite directions in the same year, and both be correct. What the behaviour surveys do and do not measure is set out in our reading of what the student AI use numbers actually measure.
The three generators of a case count
1. Detection effort
Cases begin with something being noticed. Extend screening into a department that was not screened, and its cases appear where there were none — not because behaviour changed but because coverage did. The same holds for a new tool, a new report configuration, or a marker who has just been trained to read one.
This makes the case count partly a function of your own procurement calendar, which is an uncomfortable but useful thing to know before you present a trend.
2. Definitions
Whether unauthorised AI use is a distinct category or a subtype of plagiarism changes the count. Whether a first offence is handled locally or centrally changes whether it enters the series at all. Whether contract cheating is recorded separately changes whether it is visible.
Most institutions revised their definitions between 2023 and 2026 to address generative AI. Any series spanning that revision has a break in it, and a trend drawn across the break is measuring the redefinition.
3. Reporting culture
Markers decide whether to raise a concern. That decision responds to how long the process takes, whether it is perceived as fair, whether the marker expects to be supported, and how confident they feel about the evidence. Reduce the friction and the count rises; leave a slow process in place and experienced staff route around it.
This third generator is the one nobody puts in the footnote, and it is often the largest.
The denominator problem
Case rates are usually expressed per enrolled student. That is the wrong denominator, and it fails in a specific, predictable way.
A rate per enrolled student conflates coverage with incidence. Expand screening into large low-risk assessments and the rate falls while the absolute count rises. Restrict screening to high-risk assessments and the rate climbs while nothing changes about students. Both movements are artefacts.
Use submissions screened as the denominator, and report three numbers rather than one:
- Cases opened — an institutional act, defined by your regulations.
- Submissions screened — the exposure base.
- Coverage — screened submissions as a proportion of all submissions.
Three numbers make a change interpretable. One number does not, and one number is what almost every dashboard reports.
What a case count can legitimately support
Quite a lot, as long as the claim stays within the measure:
- Workload and resourcing. Cases multiplied by median staff hours per case is a defensible operational figure and the strongest argument in most business cases.
- Process performance. Time from concern to outcome, proportion upheld, proportion appealed, appeal success rate. These describe the process and are genuinely comparable year to year within one institution.
- Distribution. Where cases arise by level, mode and assessment type, which points at assessment design rather than at students.
- Exposure modelling. Submissions multiplied by a stated tool error rate, expressed as expected cases — the arithmetic in our analysis of whether AI detection is reliable enough to base a case on.
And what it cannot support: a prevalence claim, a comparison with another institution, a statement about whether misconduct is rising, or a cross-year trend spanning a definitional change.
How to build a series worth having
Six steps, none requiring new systems.
- Write the case definition down and version it. Date every change. A series without a version history cannot be trended.
- Record coverage every term. Which assessments were screened, in which departments, with which tool. This is the single most valuable field and the most commonly missing.
- Separate the stages. Concern raised, case opened, outcome reached, appeal lodged. Collapsing them hides whether a change is in detection or in process.
- Record the trigger. Whether the concern originated from a similarity match, an authorship indicator, a marker’s own reading or a third-party report. Cases collapse on appeal when nobody can say afterwards — and the trigger field is also what makes a tool’s contribution measurable. The distinction the field encodes is in our note on how plagiarism detection and AI detection differ.
- Publish the caveats with the number, internally. Two sentences naming coverage and the last definitional change, attached to the figure wherever it appears.
- Never benchmark externally without the other institution’s definitions. If you cannot obtain them, the comparison is not available, and saying so is the correct professional answer.
Step two is worth the effort even if you do nothing else. Coverage is what converts an uninterpretable count into a measure, and it is usually already recorded somewhere in the screening tool’s own reporting.
What to do about the sector gap
The absence of a comparable cross-institutional series is unlikely to be resolved soon, because the barrier is definitional rather than technical: institutions would have to agree what counts as a case before anyone could count them together.
Two things are worth doing in the meantime. Within a mission group or a regional consortium, agreeing a minimum common definition and a coverage field is achievable at a scale where the participants can actually negotiate, and it produces the only comparable data that will exist. And internally, a stable local series with an honest denominator is worth more to your decisions than any national figure would be, because it samples your students, your assessments and your policy.
The same reasoning applies to any measurement you build alongside a platform deployment — design the instrument before the rollout, not after, using the approach in our guide to running a departmental pilot. And when a figure does appear in a paper, apply the currency test in how current your higher education data actually is.
To discuss what a defensible internal integrity baseline looks like alongside an evaluation, request an institutional evaluation.
Frequently asked questions
Are there national academic integrity violation statistics?
Not as a recurring official series comparable to enrolment or qualification data. What exists is institutional reports, access-to-information releases, vendor material and behaviour surveys.
Does a rising caseload mean misconduct is rising?
Not on its own. Detection coverage, definitions and reporting friction all move the count independently of behaviour.
Why is a rate per enrolled student wrong?
Because it conflates coverage with incidence. Expanding screening into low-risk assessments lowers the rate while raising the count.
What denominator should we use?
Submissions screened, reported alongside cases opened and coverage. Three numbers, not one.
Can we benchmark against peer institutions?
Only with their case definitions and coverage in hand. Without those, the comparison is not available.
Can survey prevalence figures substitute for case data?
No. Surveys measure self-reported behaviour; cases measure institutional action. They answer different questions.
How should we handle the AI-related definition change?
Treat it as a break in series, date it, and do not trend across it without saying so.
What single field should we start recording?
Coverage — which assessments were screened, where, with what. It is what makes the count interpretable.
Why record the trigger for each case?
Because it distinguishes concerns arising from a matched passage, an authorship score, a marker’s reading or a report, and that distinction determines what the case can rest on.
Is a falling caseload good news?
Unknown without coverage. It is equally consistent with better assessment design and with markers deciding the process is not worth the effort.
What is the strongest legitimate use of the number?
Workload. Cases multiplied by median staff hours is defensible, operationally meaningful and hard to dispute.
What should the footnote say?
Coverage, the date of the last definitional change, and the denominator used. Two sentences, attached wherever the figure travels.